Search papers, labs, and topics across Lattice.
Affiliation:
2
0
2
Ball-proximal optimization achieves super-exponential $\exp(-\Omega(K^2))$ objective decay and finite-step termination on general convex functions without requiring smoothness or strong convexity.
The first clean nonconvex convergence guarantee for pipeline parallelism reveals surprising trade-offs between methods depending on the training objective.
